/
Work with us

21st Labs

Work with 21st Labs

We're the research team at 21st. We measure how well models build and copy real, moving UI, and we can run yours before release.

What we have

1,400
Graded copies from 14 models, each with its full agent trace, one-shot and self-verified.
50
Tasks. 13 are public on GitHub, the rest are held out.
249
Copies scored by hand. Our judges match them at 0.84.
0.85
Agreement between our two judges, GPT-6.1 Sol and Claude Opus 5.5.
Harbor
Every task is packaged in Harbor format and runs on Valkyrie as copy-bench/copy-bench@v1.
Thousands
Of components in the 21st.dev catalog, a steady source of fresh tasks.

What we can do for you

Pre-release runs

We run your model through our benchmarks before launch and send you the full report.

Private evals

We test on held-out tasks that never leave our side.

Custom task sets

We build tasks from the catalog for the UI skills you care about, like motion, 3D, shaders and design systems.

Benchmarks

Copy Bench

v1

Models rebuild a web component from a few seconds of screen recording. We score how close each copy gets.

Design System Bench

v1

Coding agents build UI from a short ticket in a real product's repo. We score how often they use its own design-system components.

Flow Bench

Coming soon

Does an agent keep a product consistent once it leaves the hero screen.

21st Labs

Test your model on real UI work

  • Run your model through our benchmarks before release.
  • Private evals. Only 13 of the 50 Copy Bench tasks are in the public repo.
  • Custom task sets built from the 21st.dev component catalog.